How I write code now

Vague goal

When I know roughly where I want to go but not the way there, the AI agent and I find it together.

  1. AI agentMe

    Alignment with /wayfinder

  2. Map of clear features

Clear feature

When I know what I want, the main agent writes the plan and a second model checks it before any code is written.

  1. Main agentMe

    Alignment

  2. Main agent

    Implementation plan

    Until no blockers

    Reviewer Subagent

    Another model

  3. Coder Subagent

    Executes the plan

  4. PR

Code review

An automated loop keeps looking for PRs and reviews each in two layers, and the main agent works through the feedback.

  1. Automated reviewer agent

    General review + Security review

    Keeps looking for PRs

  2. PR

  3. Main agent

    Addresses feedback

  4. Until no blockers

    Merging

Gardening

Left alone, a codebase quickly turns into a mess, so it needs a gardener.

  1. Gardener manager agent

    Finds candidates: libraries, crates, etc.

    Keeps looking

  2. Gardener subagents

  3. Trim / deslop